Skip to main content
A scenario describes one situation to simulate. It has a stable ID, a name, and explicit input for your agent. It can also select Tool datasets, define a sequence, or request assessment. A scenario can demonstrate success, explore a boundary, reproduce a failure, or simply record behavior. Checks are optional.

​
Keep task input separate from evaluation

Put knowledge the agent should receive in input. Exact checks, scoring rubrics, dataset selection, and future sequence controls belong to the simulation definition, not the agent’s task prompt. A target receives scenario correlation identity, not a copy of the answer key.

​
Reuse and vary

Run scenarios against different configurations and repeat them to observe output variation. Save definitions in your repository and retain scenario revisions in Firedrill. Your coding agent can draft candidates from your application’s actual interfaces. Scenario recommendations help prioritize authored candidates; they do not prove complete coverage. See scenario matrices.